本文为德国汉堡大学（作者：ANDRIY BOROVKOV）的硕士论文，共55页。

本文提出了一种用于银河系图像分类器学习的可视化反卷积网络方法。在分类过程中，采用了一种具有视点提取的卷积神经网络进行预处理。该网络能够实现训练数据集的RMSE为0.084，验证数据集的RMSE为0.082。反卷积技术表明，该网络能够通过卷积层分离和增强不同的目标以获得不同的星系视角，从而突出星系的形态特征。

A deconvolutional network approach for thevisualization of the learning of a galaxy image classifier is presented here.For the classification, a convolutional neural network with viewpointextraction in preprocessing is used. The network was able to achieve a RMSE of0.084 for the training data set and 0.082 for the validation data set. Thedeconvolution technique shows that the network is able to highlight themorphological features of the galaxies through the convolutional layers byisolating and enhancing different objects for different views of a galaxy.

1 引言

2 深度学习

3 星系图像数据与Kaggle竞赛

4 设置

5 超参数优化

6 神经网络的分解

7 展望

8 结论

更多精彩文章请关注公众号：